Unlocking the Power of Inline Floating-Point Operations on Programmable Switches

Explore a groundbreaking 16-minute conference talk from USENIX NSDI '22 that introduces FPISA, a novel floating-point representation designed for programmable switches. Delve into the challenges of implementing floating-point operations on existing switch hardware and discover how FPISA addresses these limitations. Learn about the proposed hardware changes based on the open-source Banzai switch architecture and their synthesis in a 15-nm standard-cell library. Examine the performance improvements FPISA brings to accelerators for machine learning training and query processing, including reduced CPU core usage and increased throughput compared to existing solutions like SwitchML and Spark. Gain insights into the potential of FPISA to revolutionize application-level functionality acceleration in programmable dataplanes.